Two-Phase Neighbor Consensus Sharpness-Aware Minimization for Learning With Noisy Labels
TL;DR
Two-Phase Neighbor Consensus Sharpness-Aware Minimization (TNCS) is presented, which constructs soft labels from local prediction agreement without requiring an early binary separation of clean and noisy samples.
Abstract
Learning with noisy labels remains challenging because deep neural networks can eventually memorize incorrect annotations. Although Sharpness-Aware Minimization (SAM) has been reported to delay this behavior in some noisy-label settings, its optimization still relies on the observed noisy labels. This paper presents Two-Phase Neighbor Consensus Sharpness-Aware Minimization (TNCS), which constructs soft labels from local prediction agreement without requiring an early binary separation of clean and noisy samples. Phase 1 generates a fixed KNN soft-label bank from learned features and predictions. Phase 2 independently re-initializes the model and retains the observed hard label when it agrees with the Phase-1 consensus class, while using the complete soft label otherwise. TNCS obtains accuracies of 90.40% and 90.24% on Food-101 and Animal-10N, respectively. Compared with Phase-1-only training, adding Phase 2 increased mean accuracy by 4.02 and 12.82 percentage points in the two CIFAR-100 ablation conditions with 60% uniform and 40% pair-wise noise, respectively. Diagnostic analyses indicate that consensus-label quality differs among the evaluated noise conditions and that Phase-1 refined-label accuracy is positively associated with Phase-2 accuracy after condition-specific mean differences are removed in four CIFAR-100 conditions. The observed contribution of selective consensus-label transfer varies by dataset and noise condition.